Advanced Algorithms for Exploitation of Space-Based Imagery to Detect Targets against Structured Earth Backgrounds

Signature-based hyperspectral detection algorithms have demonstrated encouraging results over a range of imagery. These algorithms require as input a reflectance or emissivity spectrum for the target material of interest. Signature-based detection algorithms are often demonstrated using a carefully measured laboratory spectrum for an ideal sample of a target material. In many situations, however, a laboratory spectrum will not be available for a target material. In some cases, we may have a remotely sensed radiance spectrum obtained by an image analyst for a target or multiple field measurements from different sources. A remote spectral measurement cannot be used directly for detection in other scenes because it depends on the environmental conditions at the time of measurement. In the most challenging cases, the remotely sensed radiance spectrum may be for a subpixel target or for a target that is concealed under trees. We have shown that the performance of signature-based detection algorithms has a significant dependence on the input target spectral model. Thus, the generation of optimized target models from different input sources is a critical issue for the success of target detection systems in operational environments. There is also an important relationship between the utility of a target model and the spectral properties of the background clutter in a scene. In this project, we will develop an innovative software product called MODGEN that produces target models that optimize contrast with the background using input laboratory, field, and remote spectral measurements. This product can be used in combination with signature-based detection algorithms such as HyperTech's invariant detection products to provide a system that can be used operationally for model selection and target detection. MODGEN will consider both the visible through short-wave infrared (VNIR/SWIR) spectral range and the long-wave infrared (LWIR) spectral range. As part of this effort, we will consider the important but previously unaddressed problems of model extraction for observed targets that appear at subpixel scale or that are heavily obscured by vegetation. The problem of subpixel targets is of particular importance for space-based hyperspectral systems that often have limited spatial resolution. We will assess MODGEN by evaluating the improvement in ROC performance that is achieved using the new optimized target models as compared against the use of standard models. The assessment will consider a range of VNIR/SWIR and LWIR hyperspectral datasets.